Identifying Active Sites for CO2 Reduction on Dealloyed Gold Surfaces by Combining Machine Learning with Multiscale Simulations

Abstract

Gold nanoparticles (AuNPs) and dealloyed Au₃Fe core-shell NP surfaces have been shown to have dramatically improved performance in reducing CO₂ to CO (CO2RR), but the surface features responsible for these improvements are not known. The active sites cannot be identified with surface science experiments, and quantum mechanics (QM) is not practical for the 10 000 surface sites of a 10 nm NP (200 000 bulk atoms). Here, we combine machine learning, multiscale simulations, and QM to predict the performance (a-value) of all 5000-10 000 surface sites on AuNPs and dealloyed Au surfaces. We then identify the optimal active sites for CO2RR on dealloyed gold surfaces with dramatically reduced computational effort. This approach provides a powerful tool to visualize the catalytic activity of the whole surface. Comparing the a-value with descriptors from experiment, computation, or theory should provide new ways to guide the design of high-performance electrocatalysts for applications in clean energy conversion.